An ai-based medical examination information recognition method and system

Abstract

The application discloses an AI-based medical examination information recognition method and system, and relates to the technical field of medical image processing and character recognition. The method aims to solve the problem that the traditional OCR method uses unified preprocessing parameters for the whole image, which cannot adapt to internal quality differences, resulting in limited recognition effect. The method mainly includes: preliminary processing and recognition of medical document images, calculation of risk values by comprehensively considering the confidence of the character area and the image quality parameters, and screening of high-risk target areas; dynamically adjusting the region boundary box according to the test item rules; generating a second set of adaptive preprocessing parameters by iterative optimization for the region; enhancing the target area using the parameters, and synthesizing with the rest of the region image, and finally recognizing and outputting the results. The application can accurately locate and adaptively enhance the low-quality area in the image, effectively improve the accuracy and robustness of complex medical document character recognition, and is suitable for the automatic processing of medical documents such as test reports.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and character recognition technology, and more specifically, this application relates to an AI-based method and system for medical examination information recognition. Background Technology

[0002] With the deepening application of artificial intelligence technology in the field of medical informatics, automated processing of medical documents based on image recognition has become a key direction for improving diagnostic efficiency and data management quality. Medical test reports, medical records, and other documents often exist in image form, and their accurate recognition and structuring are of great significance for clinical decision support and medical data mining. Currently, mainstream methods typically rely on overall preprocessing of the document images followed by text extraction using a unified optical character recognition model, which reduces the algorithmic complexity and computational resource consumption.

[0003] However, these methods implicitly assume that the quality distribution of the entire image is relatively uniform, thus employing fixed preprocessing parameters to enhance the overall readability of the image. However, in real-world medical scenarios, due to factors such as shooting conditions, paper material, printing quality, or storage wear and tear, regional quality differences often exist within the same medical document image; for example, some characters may be blurry, have low contrast, or be subject to smudges or other interference.

[0004] If a globally uniform preprocessing strategy is continued, enhancing images to improve the recognition of low-quality areas may overprocess originally clear, high-quality areas, introducing noise or distortion. Conversely, if the parameter settings are too conservative, it will be difficult to effectively improve the recognition bottleneck of low-quality areas. This "one-size-fits-all" approach essentially prevents the preprocessing stage from adapting to the non-uniform characteristics of image content, causing the subsequent recognition model to face an inherent ceiling in improving its overall accuracy when dealing with complex real-world scenarios. Therefore, an AI-based medical laboratory information recognition method and system are proposed to address this problem. Summary of the Invention

[0005] To address the aforementioned technical problems, this paper provides an AI-based method and system for identifying medical laboratory information. This technical solution resolves the issues raised in the background section.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] Firstly, this application provides an AI-based method for identifying medical test information, including:

[0008] The medical document image to be processed is acquired, and preprocessed using a first set of preprocessing parameters to obtain a first intermediate image;

[0009] Perform character recognition on the first intermediate image to generate an initial recognition result including multiple character regions and their comprehensive risk quantification values, and take the character regions with comprehensive risk quantification values ​​greater than or equal to a preset risk threshold as the target optimization regions;

[0010] Obtain the initial bounding box of the target optimization region and the identifiers of each item;

[0011] Query the pre-stored rule table to obtain the standard character length and unit symbol position rules corresponding to the project identifier. Based on this, calculate the width adjustment amount and adjustment direction of the initial bounding box, and adjust it to obtain the adjusted bounding box.

[0012] If the width of the adjusted bounding box is less than the preset width threshold, then adjust its width to the width threshold while keeping the center unchanged, and output the area defined by the adjusted bounding box as the final processing area.

[0013] Based on the medical document image fragments of the final processing region, iterative optimization is performed within a preset parameter search space to generate a matching second preprocessing parameter set; wherein, the second preprocessing parameter set is different from the first preprocessing parameter set;

[0014] The final processing region is enhanced using the second preprocessing parameter set, while the remaining regions are processed using the first preprocessing parameter set, and then combined into a second intermediate image.

[0015] Character recognition is performed on the final processing region in the second intermediate image to obtain optimized results, and the final recognition result is generated by combining the initial recognition results of the remaining regions.

[0016] Secondly, this application provides an AI-based medical laboratory information recognition system for implementing the aforementioned AI-based medical laboratory information recognition method, comprising:

[0017] The image acquisition module is used to acquire the image of the medical document to be processed, and preprocess it using a first set of preprocessing parameters to obtain a first intermediate image;

[0018] The region filtering module is used to perform character recognition on the first intermediate image, generate an initial recognition result including multiple character regions and their comprehensive risk quantification values, and take the character regions with comprehensive risk quantification values ​​greater than or equal to a preset risk threshold as target optimization regions.

[0019] The boundary adjustment module is used to obtain the initial bounding box of the target optimization area and the identifiers of each item, and to query the pre-stored rule table to obtain the standard character length and unit symbol position rules corresponding to the item identifiers. Based on this, the width adjustment amount and adjustment direction of the initial bounding box are calculated, and the adjusted bounding box is obtained.

[0020] The width adjustment module is used to adjust the width of the bounding box to the preset width threshold if the width of the bounding box is less than the preset width threshold, while keeping the center unchanged, and output the area defined by the adjusted bounding box as the final processing area.

[0021] The parameter optimization module is used to iteratively optimize the medical document image fragments based on the final processing region within a preset parameter search space to generate a matching second preprocessing parameter set; wherein the second preprocessing parameter set is different from the first preprocessing parameter set;

[0022] The image enhancement module is used to enhance the final processing area using the second preprocessing parameter set, while processing the remaining areas using the first preprocessing parameter set, and synthesizing them into a second intermediate image.

[0023] The result output module is used to perform character recognition on the final processing area in the second intermediate image to obtain optimized results, and combine the initial recognition results of the remaining areas to generate the final recognition result.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0025] This application dynamically adjusts the size and orientation of the target region's bounding box based on the standard character length and unit symbol position rules corresponding to the project identifier. This solves the problem of inaccurate region truncation caused by inconsistent character lengths and variable unit symbol positions in medical testing projects, enabling complete and accurate extraction of test items and result values, and providing reliable region input for subsequent classification and recognition.

[0026] This application generates a matching second preprocessing parameter set for the final processing region by iteratively optimizing within a preset parameter search space, and then performs image synthesis by applying differentiated preprocessing parameters to different regions. This solves the limitation that a single global preprocessing parameter cannot take into account local image features, achieves adaptive enhancement of local image quality in medical documents, improves the accuracy of character recognition in complex backgrounds, low-contrast or blurred areas, and ultimately forms a reliable structured recognition result. Attached Figure Description

[0027] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Wherein:

[0028] Figure 1 This is a flowchart of an AI-based medical test information recognition method proposed in this invention;

[0029] Figure 2 This is a data flow diagram of the AI-based medical test information recognition method in this invention;

[0030] Figure 3This is a schematic diagram of the structure of an AI-based medical testing information recognition system proposed in this invention. Detailed Implementation

[0031] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0032] In traditional image recognition-based automated medical document processing methods, the recognition process typically relies on applying a uniform set of preprocessing parameters to the entire image. The problem with this approach is its inability to adapt to the non-uniform quality of different regions within the image. When medical document images suffer from blurred characters, low contrast, or stains due to factors such as shooting conditions, paper wear, or printing quality, globally fixed preprocessing parameters cannot simultaneously optimize the recognition performance of high-quality and low-quality areas. Enhancing the image to improve the recognition of low-quality areas may introduce noise or distortion into clear areas; conversely, conservative parameter settings make it difficult to improve the recognition bottleneck in low-quality areas. This "one-size-fits-all" approach limits the overall accuracy improvement of the preprocessing stage and subsequent recognition models when facing complex real-world scenarios.

[0033] For example, in a blood test report scanned in a hospital's laboratory, the patient's name and test item names were clearly printed with a clean background. However, the contrast of some test result numerical areas was significantly reduced due to fading of the printing ribbon. Traditional systems use uniform grayscale stretching and binarization parameters to process the entire image. While enhancing the contrast of the numerical areas, this results in overexposure and stroke blurring in the name and test item name areas. The character recognition model's confidence level for high-quality areas decreased from 0.95 to 0.82, while the confidence level for low-quality numerical areas only increased from 0.45 to 0.60, leading to an overall increase in the recognition error rate.

[0034] If the aforementioned problems are not addressed, automated identification systems will be unable to reliably handle documents of varying quality commonly encountered in real-world medical scenarios, leading to missed or misidentified critical testing information and impacting the accuracy and efficiency of clinical decision-making. Furthermore, the system's deployment is subject to stringent image quality requirements, exhibits poor generalization ability, necessitates frequent manual correction, and increases operational costs. This mismatch between preprocessing strategies and local image features severely restricts the widespread application of automated medical document processing technology in real-world environments.

[0035] To address the aforementioned issues, this application first analyzes the root cause of the failure of traditional methods, finding that the problem lies in the failure to consider regional quality differences and recognition risks in the preprocessing stage. In response, this application attempts to accurately locate and identify bottleneck regions based on comprehensive risk quantification values. Furthermore, this application proposes dynamically adjusting the bounding box of the target region based on project identification rules to ensure complete extraction of key information; and iteratively optimizing high-risk regions within a preset parameter search space to generate an adaptive second preprocessing parameter set. After verifying the multi-region differentiated processing strategy, it is finally determined to adopt a combination of risk quantification screening, bounding box rule adjustment, and local parameter optimization to enhance high-risk regions while retaining the original processing results of other regions, synthesizing them into an optimized image, thereby achieving high-precision and robust recognition of medical test information.

[0036] Example 1:

[0037] like Figure 1-2 As shown, this application introduces an AI-based method for identifying medical test information, including:

[0038] S1. Obtain the image of the medical document to be processed, and preprocess it using the first set of preprocessing parameters to obtain the first intermediate image;

[0039] Receive images of medical documents to be processed. These images can come from the digital document library of a medical information system or be acquired through devices such as scanners. Common formats include JPEG and PNG. The image content is usually a test report or medical record containing text and tables.

[0040] The acquired image is standardized using a preset first preprocessing parameter set. This first preprocessing parameter set contains a set of fixed parameters applicable to most scenarios. Typical processing steps include: grayscale conversion: converting the color image to a grayscale image; noise reduction: reducing image noise using methods such as mean square filtering; contrast enhancement: improving the overall contrast of the image through histogram equalization; binarization: converting the grayscale image to a black and white binary image using a global threshold (e.g., automatically determined by the Otsu algorithm) to separate the foreground text from the background; and tilt correction: detecting and correcting slight page tilt based on image content.

[0041] The above operations are performed sequentially and work together on the entire input image.

[0042] After the above processing, a first intermediate image is generated and output. This image is a binary or grayscale image that has undergone preliminary enhancement.

[0043] Example: Take a scanned blood test report image (PNG format) as an example. After acquiring the image, the system applies the first set of preprocessing parameters in sequence: first, grayscale conversion and 3×3 median filtering for noise reduction; then, histogram equalization to enhance contrast; next, binarization using the global threshold calculated by Otsu's algorithm; finally, detection and rotation correction of a 0.5° tilt. In the final output first intermediate image, the text and background have been initially separated, and the overall clarity has been improved, but regional quality differences may still exist within the image.

[0044] Through the above technical solution, this application solves the problem of insufficient adaptability caused by fixed preprocessing parameters in medical document image processing. It dynamically adjusts the preprocessing strategy for different types and qualities of original images, which not only avoids feature loss caused by overprocessing of low-quality images, but also ensures that high-quality images can fully retain key information. At the same time, through intelligent adaptation of the first preprocessing parameter set, it improves the accuracy and processing efficiency of subsequent image analysis, making the digitization and recognition results of medical documents more in line with the actual clinical application needs.

[0045] S2. Perform character recognition on the first intermediate image to generate an initial recognition result including multiple character regions and their comprehensive risk quantification values, and take the character regions with comprehensive risk quantification values ​​greater than or equal to the preset risk threshold as the target optimization regions.

[0046] The process of obtaining the comprehensive risk quantification value specifically includes:

[0047] Obtain the confidence level of each character region, perform image quality assessment on each character region, and generate corresponding quality assessment parameters;

[0048] For each character region, a comprehensive risk quantification value is obtained by weighting its confidence level and quality assessment parameters.

[0049] The formula for calculating the comprehensive risk quantification value is as follows:

[0050] ;

[0051] in, This represents the overall risk quantification value. This represents the normalized confidence level. This represents the normalized quality assessment parameters. , Represents the weighting coefficient, and .

[0052] Obtain the initial recognition results, which include the confidence score C and quality assessment parameter Q for each character region.

[0053] Calculate the comprehensive risk quantification value R: For each character region, according to the formula... Calculate its risk value. , Preset weighting coefficients ( This is used to balance the impact of recognition uncertainty and image quality.

[0054] Target optimization region selection: The R value of each region is compared with a preset risk threshold. If R is greater than or equal to the threshold, the region is determined to be a high-risk region. Subsequently, multiple high-risk regions that are spatially adjacent can be merged to form one or more contiguous target optimization regions.

[0055] Output information about one or more target optimization regions, including their merged bounding box coordinates and corresponding project identifiers.

[0056] Example: For a character region with confidence level C=0.25 and quality parameter Q=0.30, take... The calculated R value is 0.725. If the preset risk threshold is 0.6, this area is marked as high-risk. The system merges it with other high-risk characters belonging to the "HGB: 12.5g / dL" item, and finally outputs a target optimization area that covers the "12.5" value portion.

[0057] Through the above technical solution, this application solves the problem of inaccurate optimization region selection caused by a single indicator judgment in the post-processing of medical document image recognition. For each character region that has been identified, the comprehensive risk quantification value is dynamically calculated by normalizing and weighting the confidence level and image quality assessment parameters. This avoids the mis-optimization caused by relying solely on low confidence level and ignoring the image clarity itself, and also prevents subsequent errors caused by the masking of recognition uncertainty due to a single quality assessment.

[0058] The calculation process for the confidence score of each character region specifically includes:

[0059] Obtain the candidate character set corresponding to each character region generated by the character recognition process, and the confidence score of each candidate character in the candidate character set;

[0060] Extract the top M candidate characters with the highest confidence from the candidate character set as Top-M candidate characters, and obtain their corresponding confidence scores, where M is a preset positive integer;

[0061] The confidence scores of the Top-M candidate characters are normalized to obtain the normalized confidence score of each candidate character.

[0062] Based on the normalized confidence level, the information entropy value of the character region is calculated using the following formula:

[0063] ;

[0064] in, This represents the information entropy value, and j represents the index of the candidate character. Represents the normalized confidence score of the j-th Top-M candidate character;

[0065] The ratio of the information entropy value to the preset baseline entropy value is used as the confidence level of the character region.

[0066] The process of obtaining the above confidence levels and quality assessment parameters specifically includes:

[0067] Obtain the first intermediate image output from step S1.

[0068] The following two processes are executed in parallel:

[0069] Character Recognition and Confidence Generation: An OCR engine (e.g., a CRNN model combining convolutional and recurrent neural networks, or a Transformer-based OCR model) is used to recognize character regions in the image, outputting a set of candidate characters and their probabilities for each region. For each region, the top M (e.g., M=3) candidate characters with the highest confidence are selected from the candidate set, and their probabilities are normalized to obtain... Then, according to the formula Calculate the information entropy of this region. Finally, the normalized confidence level C of the character region was calculated using the min-max method based on historical data.

[0070] Image quality assessment: For each character region, the image block is calculated to evaluate its sharpness, contrast, noise level, and other indicators. These are then weighted and combined to generate a normalized quality assessment parameter Q (a higher Q value indicates better image quality).

[0071] Generate structured initial recognition results. This result is a list, where each entry corresponds to a character region, including its bounding box coordinates, the most likely character identified, the calculated confidence score C, and the quality assessment parameter Q.

[0072] Example: Take a region in the first intermediate image that is identified as the digit "5" as an example. The OCR engine provides a candidate set {'5': 0.88, 'S': 0.07, '8': 0.03}. Taking M=3, after normalizing the probabilities, its information entropy H is calculated to be approximately 0.57, thus yielding a confidence level C of approximately 0.36. Simultaneously, analyzing the image patch of this region yields a quality assessment parameter Q=0.45. The result for this region is then recorded in the initial recognition result list.

[0073] Through the above technical solution, this application solves the problem of low reliability of recognition results caused by the rigidity of traditional confidence calculation methods in medical document image character recognition. It performs dynamic confidence assessment and quality parameter generation for regions with different image quality and character features, which not only prevents misjudgment or recognition omissions caused by overly simplified confidence calculation in low-quality image regions, but also provides a refined confidence measurement based on the information entropy of multiple candidate characters for high-definition character regions.

[0074] The weighting coefficients of the comprehensive risk quantification value are dynamically adjusted, specifically including:

[0075] Calculate the absolute difference between the confidence level and the quality assessment parameter as the parameter variability;

[0076] If the parameter difference is greater than or equal to the preset contradiction threshold, check whether there are predefined types of image interference features in the medical document image fragments corresponding to the character region;

[0077] If it exists, the quality assessment parameter is assigned a weight coefficient greater than the confidence level;

[0078] If it does not exist, the confidence level is assigned a weight coefficient higher than that of the quality assessment parameter;

[0079] The comprehensive risk quantification value is obtained by weighting the confidence level and quality assessment parameters based on the adjusted weighting coefficients.

[0080] Specifically, calculate the parameter difference D = |C−Q| for this character region.

[0081] If D is greater than or equal to the preset inconsistency threshold (e.g., 0.3), then further check whether there are predefined interference features (e.g., obvious stains, background stripes) in the original image fragment of the region.

[0082] Among them, the predefined interference features include: blur features: detected by calculating the average value of the image gradient amplitude, and judged as blurry if it is lower than the preset gradient threshold; stain features: identified by the color abnormality region detection algorithm; shadow features: identified by the analysis of brightness distribution non-uniformity; and reflective features: detected by the highlight region identification algorithm.

[0083] If interfering features are present, then the image quality issue is determined to be the primary problem, and the image quality should be increased. Set the value (e.g., to 0.7) and lower it accordingly. (Set to 0.3) to make quality risk the dominant factor in the calculation.

[0084] If no interfering features are present, then the uncertainty of the recognition model is determined to be the primary problem, and the adjustment is increased. Set the value (e.g., to 0.7) and lower it accordingly. (Set to 0.3).

[0085] Use the adjusted and Recalculate the R value.

[0086] Through the above technical solution, this application solves the problem of inaccurate calculation of comprehensive risk quantification value caused by fixed weight allocation in medical document image processing. It dynamically adjusts the weight coefficients for the difference between the confidence level and quality assessment parameters of different character regions, which not only prevents the comprehensive assessment distortion caused by the rigidity of weights when there are significant contradictions in the parameters, but also provides an adaptive weight allocation mechanism for regions with image interference or recognition uncertainty, ensuring that risk quantification can flexibly adapt to complex and changing image conditions.

[0087] The risk threshold is dynamically adjusted, specifically including:

[0088] Obtain image quality evaluation parameters for all character regions;

[0089] Calculate the global mean and global standard deviation of the image quality assessment parameters;

[0090] Based on the combination of the global average and the global standard deviation, the preset risk threshold is adjusted to generate the adjusted risk threshold.

[0091] The specific rules for adjusting based on the combined state are as follows:

[0092] If the global average is lower than the quality benchmark threshold and the global standard deviation is lower than the uniformity benchmark threshold, the preset risk threshold will be increased.

[0093] If the global average is higher than the quality benchmark threshold and the global standard deviation is higher than the uniformity benchmark threshold, then the preset risk threshold is lowered.

[0094] If the global average is higher than the quality benchmark threshold and the global standard deviation is lower than the uniformity benchmark threshold, then the preset risk threshold is increased.

[0095] If the global average is lower than the quality benchmark threshold and the global standard deviation is higher than the uniformity benchmark threshold, then the preset risk threshold is lowered.

[0096] This section is used to dynamically adjust the risk threshold in step S3 based on the overall image quality, thereby improving the screening adaptability. Its specific implementation method is as follows:

[0097] Obtain the quality assessment parameter Q for all character regions of the current image (derived from the output of step S2).

[0098] Calculate statistics: Based on the input set of Q values, calculate the global mean and global standard deviation.

[0099] Judgment and Adjustment: Compare the global average with the preset quality benchmark threshold, and compare the global standard deviation with the preset uniformity benchmark threshold. Based on the comparison results, adjust the original preset risk threshold according to the following rules:

[0100] If the global average is less than the quality benchmark threshold and the global standard deviation is less than the homogeneity benchmark threshold, then the original preset risk threshold is increased (e.g., the original threshold is multiplied by a coefficient greater than 1.25).

[0101] If the global average is greater than the quality benchmark threshold and the global standard deviation is greater than the homogeneity benchmark threshold, then the original preset risk threshold is lowered (e.g., the original threshold is multiplied by a coefficient less than 0.75).

[0102] If the global average is greater than the quality benchmark threshold and the global standard deviation is less than the homogeneity benchmark threshold, then the original preset risk threshold is increased (e.g., the original threshold is multiplied by a coefficient greater than 1 but less than 1.25).

[0103] If the global average is less than the quality benchmark threshold and the global standard deviation is greater than the homogeneity benchmark threshold, then the original preset risk threshold is lowered (e.g., the original threshold is multiplied by a coefficient less than 1 but greater than 0.75).

[0104] The adjusted risk threshold is used to replace the original threshold for region screening in step S3.

[0105] Example: The Q-values ​​for all character regions in an image are: [0.85, 0.82, 0.30, 0.88, 0.25, 0.90, 0.28, 0.86, 0.83, 0.29]. The calculated global average is approximately 0.586, and the global standard deviation is approximately 0.290. Let the quality benchmark threshold be 0.6 and the uniformity benchmark threshold be 0.2. At this point, the global average is less than the quality benchmark threshold and the global standard deviation is greater than the uniformity benchmark threshold, which meets the adjustment rule. If the original preset risk threshold is 0.6, and the adjustment coefficient is 0.9, then the adjusted threshold is 0.54. 0.54 will be used as the new screening threshold subsequently.

[0106] Through the above technical solution, this application solves the problem of insufficient adaptability of optimization region screening caused by fixed risk threshold in medical document image processing. The risk judgment threshold is dynamically adjusted according to the overall quality distribution and uniformity of different images. This avoids optimization omissions caused by overly strict thresholds when the overall image quality is generally low and the distribution is uniform, and also prevents unnecessary processing burden caused by overly wide thresholds when the overall quality is high but local differences are significant.

[0107] S3. Obtain the initial bounding box of the target optimization region and the identifiers of each item;

[0108] Query the pre-stored rule table to obtain the standard character length and unit symbol position rules corresponding to the project identifier. Based on this, calculate the width adjustment amount and adjustment direction of the initial bounding box, and adjust it to obtain the adjusted bounding box.

[0109] The process of calculating the width adjustment amount and direction of the initial bounding box specifically includes:

[0110] Based on the standard character length and unit symbol position rules, the theoretical total text length of the target optimization region is calculated;

[0111] Based on the width of the initial bounding box and by analyzing the arrangement of each character region within it, the actual character region width of the target optimization region is obtained.

[0112] The theoretical total length of the text is compared with the actual width of the character area, and the difference between the two is used as the width adjustment amount; the adjustment direction is determined according to the preset orientation of the unit symbol relative to the character, as indicated by the unit symbol position rules.

[0113] Regarding step S3, it includes:

[0114] Obtain the target optimization region information (including initial bounding box coordinates and project identifier) ​​output from step S2, as well as a pre-stored rule table (associating the project identifier with its standard character length and unit symbol position rules).

[0115] Query rules: Query the rule table based on the project identifier to obtain the corresponding standard character length and unit symbol position rules.

[0116] Calculate the adjustment amount: Based on the obtained rules, combined with the preset average pixel width of a single character and the standard width of a unit symbol, calculate the theoretical total length of the text in the target region. Simultaneously, calculate the actual width of the character region based on the actual distribution of the identified characters within the initial bounding box. The difference between the two is used as the width adjustment amount.

[0117] Determine the direction: Determine the adjustment direction of the bounding box (e.g., expand to the right) according to the unit symbol position rules (e.g., "right side").

[0118] Generate the adjusted bounding box: Based on the calculated width adjustment amount and adjustment direction, translate the corresponding edge of the initial bounding box to obtain the adjusted bounding box.

[0119] Output the adjusted bounding box coordinates after rule calibration.

[0120] Example: The target optimization area is labeled "HGB", and its initial bounding box contains content with a width of 50 pixels. Looking up the rule table, the standard character length for "HGB" is 4, with the unit on the right and a standard unit width of 30 pixels. Assuming an average character width of 12 pixels, the theoretical total length = 412 + 30 = 78 pixels. The width adjustment ΔW = 78 - 50 = 28 pixels. Because the unit is on the right, the right edge of the bounding box is extended 28 pixels to the right to obtain the new adjusted bounding box.

[0121] The process of building a pre-stored rule table includes:

[0122] The source data includes: historical structured test report data (including items, result values, and units) from the hospital information system; and a medical document image library with precise manual annotations (including bounding boxes and text content for item result areas).

[0123] Data processing (rule extraction): For each test item:

[0124] Define the project identifier (e.g., "HGB").

[0125] Calculate the standard character length: Analyze a large number of historical result value samples, count the string length of the numerical part (excluding units), and take the mode or high frequency length as the standard character length of the project.

[0126] Determine the rules for the position of unit symbols: Analyze the orientation of unit symbols relative to numerical values ​​in the labeled image (e.g., "right adjacent"), and record the most frequent orientation patterns as rules.

[0127] Estimate standard width of unit symbols: Measure and calculate the average pixel width of commonly used units in the image as the standard pixel width.

[0128] The processing results are organized into a structured rule table and persisted. Each record contains: "Project Identifier", "Standard Character Length", "Unit Symbol Position Rule", and "Standard Pixel Width of Unit Symbol".

[0129] Rule table maintenance and updates: Initial construction is completed before system deployment; existing rules can be optimized and updated regularly based on feedback cases during operation; when encountering new projects, manual entry or a sample-based learning process can be started, and new records can be added after review.

[0130] Example: Take the "Hemoglobin (HGB)" item as an example.

[0131] Analyzing historical data, the mode of its numerical length is 5, therefore the standard character length is set to 5.

[0132] Analyzing the labeled images, the unit "g / dL" is "right adjacent" to the value in most cases, so the rule for the position of the unit symbol is set to "right adjacent".

[0133] The average width of “g / dL” in the measurement image is 28 pixels, so the standard pixel width of the unit symbol is set to 28 (or rounded to 30).

[0134] Generate a rule table record: "{"HGB", 5, "Right Adjacent", 30}.

[0135] Through the above technical solution, this application solves the problem of insufficient width adaptability caused by fixed bounding box rules in medical document image processing. It dynamically adjusts the bounding box size for the character area corresponding to different project identifiers, which not only prevents the reduction in recognition accuracy caused by the truncation of key characters due to the bounding box being too narrow or the introduction of background noise due to the bounding box being too wide, but also provides structured positioning support for medical text containing special formats such as unit symbols, ensuring that the framing process matches the actual layout of the document content.

[0136] S4. If the width of the adjusted bounding box is less than the preset width threshold, adjust its width to the width threshold while keeping the center unchanged, and output the area defined by the adjusted bounding box as the final processing area.

[0137] Regarding step S4, it includes:

[0138] Obtain the adjusted bounding box output from step S3 (based on coordinates) definition).

[0139] Calculate current width: Based on the coordinates of the adjusted bounding box, calculate its current width. .

[0140] Width determination and adjustment: With a preset width threshold Compare them. It is an empirical value used to ensure that subsequent processing has enough image information (for example, it can be set to twice the average width of a single character).

[0141] If judgment Less than Then perform the following adjustments to generate the bounding box of the final processed region:

[0142] Calculate the coordinates of the center point of the adjusted bounding box. , ),in , .

[0143] Calculate width adjustment amount .

[0144] With the center point remaining unchanged, expand the bounding box outwards on both the left and right sides simultaneously. / 2, that is, the new , . and It remains unchanged.

[0145] Output the bounding box coordinates of the final processed region. This region will be used for parameter optimization in step S6 and image enhancement in step S7.

[0146] Example: Set the preset width threshold to 50 pixels. The S4 output adjustment bounding box coordinates are [100, 200, 130, 220], and its current width is 30 pixels. Since 30 is less than 50, adjustment is triggered.

[0147] Calculate the center point =(100+130) / 2=115, =(200+220) / 2=210.

[0148] Calculate width adjustment amount =50-30=20 pixels.

[0149] Keeping the center unchanged, expand left and right by 10 pixels each, resulting in the final processing region bounding box coordinates of [115-25, 200, 115+25, 220], or [90, 200, 140, 220], with a width of 50 pixels. This region will be output.

[0150] Through the above technical solution, this application solves the problem of incomplete region coverage in medical document image processing where the bounding box may still be incomplete due to its small width after dynamic adjustment. After adjusting the bounding box according to the theoretical text length and unit symbol rules, the system further introduces a preset width threshold as a final guarantee. For adjustments where the width is insufficient, a proportional expansion is performed while keeping the center unchanged. This mechanism avoids the inability to fully contain effective character information or necessary contextual features due to the region size being too small, and also maintains the spatial consistency between the region and the original image structure by keeping the center position unchanged.

[0151] S5. Based on the medical document image fragments in the final processing region, iterative optimization is performed within the preset parameter search space to generate a matching second preprocessing parameter set; wherein, the second preprocessing parameter set is different from the first preprocessing parameter set;

[0152] The process of generating the second preprocessing parameter set specifically includes:

[0153] Obtain the medical document image fragments corresponding to the final processing area and the preset parameter search space; wherein, the parameter search space includes multiple combinations of preprocessing parameters to be evaluated;

[0154] A preprocessed parameter combination is randomly selected from the preset parameter search space as a candidate parameter combination;

[0155] Preprocessing medical document image fragments using candidate parameter combinations yields corresponding candidate processed images;

[0156] Image quality assessment is performed on candidate processed images to generate corresponding candidate assessment index values;

[0157] Based on the comparison results between the candidate evaluation index values ​​and the optimal evaluation index values ​​recorded in the historical iteration process, the optimal evaluation index values ​​and their corresponding optimal parameter combinations are updated.

[0158] Repeat this process until the iteration termination condition is met, and output the recorded optimal parameter combination as the second preprocessed parameter set.

[0159] Regarding step S5, it includes:

[0160] Input includes:

[0161] Medical document image fragment: The corresponding region image is cropped from the original medical document image (input of step S1) based on the bounding box coordinates of the final processed region output by step S4.

[0162] Predefined parameter search space: a predefined set of possible parameter combinations. For example, for the binarization step, this space may contain different threshold ranges (e.g., 120-180, step size 10); for contrast enhancement, it may contain different gamma values ​​(e.g., 0.5, 0.8, 1.0, 1.2, 1.5).

[0163] Data processing (iterative optimization):

[0164] Initialization: Set the current optimal evaluation index value to a range value (such as 0), and set the corresponding optimal parameter combination to empty.

[0165] Iterative loop: Select a candidate parameter combination sequentially or according to a strategy (such as random or grid) from the preset parameter search space.

[0166] Preprocessing and Evaluation: The medical document image fragments are preprocessed using candidate parameter combinations to obtain candidate processed images. Subsequently, the image quality of the candidate processed images is evaluated (using evaluation methods similar to those in S2, such as calculating sharpness and contrast), generating a comprehensive candidate evaluation index value (a higher value generally indicates better quality).

[0167] Comparison and Update: Compare the candidate evaluation index value with the currently recorded optimal evaluation index value. If the candidate evaluation index value is better (e.g., higher), then update the optimal evaluation index value with the candidate evaluation index value, and update the optimal parameter combination with the candidate parameter combination.

[0168] Termination Judgment: Repeat the above "selection-processing-evaluation-comparison" process until the iteration termination condition is met. This condition may be: traversing all combinations of parameters in the search space, reaching the preset maximum number of iterations, or the optimal evaluation index value improving by less than a certain threshold in a number of consecutive iterations.

[0169] When the iteration terminates, the recorded optimal parameter combination is output as the second preprocessed parameter set.

[0170] Example: Assume the final processed image segment is a low-contrast numerical region. The preset parameter search space for "contrast enhancement" includes gamma values: {0.5, 0.7, 0.9, 1.1, 1.3}. The iterative process tests sequentially:

[0171] When gamma = 0.5, the evaluation index value = 0.65, recorded as the current best; when gamma = 0.7, the evaluation index value = 0.78, better than 0.65, the best is updated to gamma = 0.7, and the best evaluation index value = 0.78; when gamma = 0.9, the evaluation index value = 0.82, the best is updated to gamma = 0.9, and the best evaluation index value = 0.82; when gamma = 1.1, the evaluation index value = 0.80, not better than 0.82, no update; when gamma = 1.3, the evaluation index value = 0.75, not better than 0.82, no update; the iteration is complete and terminated. The final output second preprocessing parameter set is: {“Contrast Enhancement_Gamma Value”: 0.9}. This parameter set will be specifically used for enhancement processing in this area.

[0172] The specific implementation methods for constructing the preset parameter search space include:

[0173] Data acquisition: The first set of preprocessing parameters used in step S1 is used as a benchmark; combined with historical experience or domain knowledge, the effective range of processing parameters for typical quality problems of medical document images (such as low contrast, blur) is considered.

[0174] Data processing:

[0175] Determine optimization parameters: Select the key preprocessing operations and their parameters that need to be optimized in S5, such as "binarization threshold" and "contrast enhancement gamma value".

[0176] Define the value range: For each selected parameter, determine a reasonable adjustment range based on its value in the first preprocessing parameter set and historical experience.

[0177] Discretization to generate candidate set: Discretize the continuous value range of each parameter according to a preset step size to generate a list of candidate values ​​for that parameter.

[0178] Combining to generate the search space: Combining all candidate values ​​of all parameters to generate the final preset parameter search space, which is the set of all parameter combinations to be evaluated.

[0179] Output a structured preset parameter search space. This space is a list where each element is a complete combination of parameters (e.g., {"binarization threshold": 140, "gamma value": 0.9}).

[0180] Example: The binarization threshold in the first preprocessing parameter set is 160, and the gamma value is 1.0. Based on experience, the binarization threshold optimization range is set to [140, 180] with a step size of 20; the gamma value optimization range is set to [0.7, 1.3] with a step size of 0.2.

[0181] Generate candidate set: threshold {140, 160, 180}, gamma value {0.7, 0.9, 1.1, 1.3}.

[0182] The combination yields a search space containing 12 parameter combinations, such as: {threshold: 140, gamma: 0.7}, {threshold: 140, gamma: 0.9}...{threshold: 180, gamma: 1.3}.

[0183] Through the above technical solution, this application solves the problem of poor local processing effect caused by globally fixed preprocessing parameters in medical document image processing. For the selected final processing area, iterative optimization is performed in the preset parameter search space to generate a suitable second preprocessing parameter set. This avoids the information loss or noise residue that may be caused by mechanically applying a single preprocessing strategy to all areas, and ensures that key areas obtain the best image enhancement effect through dynamic parameter adjustment.

[0184] During the iterative optimization process, the parameter search space is dynamically adjusted, specifically including:

[0185] During the iterative optimization process, the loss function value corresponding to each iteration is acquired and recorded in real time to form a loss function change curve.

[0186] Calculate the rate of descent and oscillation amplitude of the loss function curve;

[0187] If the rate of descent is lower than the preset rate threshold and the amplitude of oscillation is lower than the preset amplitude threshold, the optimization process is determined to be smoothly converged, and a shrinkage operation on the parameter search space is performed.

[0188] If the oscillation amplitude exceeds the preset amplitude threshold, it is determined that the optimization process is oscillating violently, and an expansion operation on the parameter search space is performed.

[0189] Based on the adjusted parameter search space, the iterative optimization process continues until the second preprocessed parameter set is generated.

[0190] The formula for calculating the value of the loss function is:

[0191] ;

[0192] in, This represents the numerical value of the loss function. This indicates the number of image quality assessment metrics. Indicates the index number. Indicates the first The target value of each evaluation indicator Indicates the first The current value of each evaluation indicator. Indicates the first The weighting coefficients of the evaluation indicators satisfy the following: Greater than or equal to 0. Used to adjust the importance of different evaluation metrics in the overall loss, satisfying... and .

[0193] In the iterative optimization of step S5, this section dynamically adjusts the parameter search space based on the state of the optimization process to improve efficiency. The specific implementation is as follows:

[0194] Dynamic input: input according to the formula in each iteration. The calculated loss function value L. Wherein, This originates from the evaluation of candidate images in step S6. and The weighting coefficients can be the same or set according to time.

[0195] Preset thresholds: Used to determine the descent rate threshold and oscillation amplitude threshold.

[0196] Recording and Calculation: Record the L value for each iteration to form a sequence. Calculate the average rate of decline V and standard deviation of this sequence within the recent window as the oscillation amplitude A.

[0197] Judgment and Adjustment:

[0198] If the average descent rate is less than the descent rate threshold and the oscillation amplitude is less than the oscillation amplitude threshold, it is determined to be a smooth convergence. Then, taking the current optimal parameter combination as the center, the search range of each parameter is reduced proportionally, and a new, more refined search space is generated by re-discretizing.

[0199] If the oscillation amplitude exceeds the oscillation amplitude threshold, it is determined to be a violent oscillation. Then, the search range of each parameter is expanded by a preset ratio to generate a new and broader search space.

[0200] If the above conditions are not met, the current search space will remain unchanged.

[0201] Output the adjusted parameter search space. The iterative optimization process will continue based on this new space until a second preprocessed parameter set is generated.

[0202] Example: Suppose the loss value sequence of the last 5 iterations is [0.50, 0.48, 0.49, 0.47, 0.475]. The calculated average descent rate V is approximately 0.005, and the oscillation amplitude A is approximately 0.012. The preset descent rate threshold is 0.01, and the oscillation amplitude threshold is 0.02. Since the average descent rate is less than the descent rate threshold and the oscillation amplitude is less than the oscillation amplitude threshold, a contraction operation is triggered. If the current optimal binarization threshold is 165, and the original search range is [140, 180], then the range is contracted to [155, 175] centered at 165, and re-discretized (e.g., with a step size of 5), generating a new threshold candidate set {155, 160, 165, 170, 175}, which serves as the new space for subsequent optimization.

[0203] Through the above technical solution, this application solves the problems of low convergence efficiency and easy getting trapped in local optima caused by fixed parameter search space in the optimization process of medical document image preprocessing parameters. The search space range is dynamically adjusted according to the changes in the loss function during the iteration process, which not only avoids unnecessary computational losses caused by continuous large-scale search when the optimization trajectory is flat, but also prevents the optimization process from being unstable due to the limitation of search space when the update oscillates violently.

[0204] S6. Enhance the final processing area using the second preprocessing parameter set, while processing the remaining areas using the first preprocessing parameter set, and synthesize them into a second intermediate image.

[0205] The process of synthesizing the second intermediate image specifically includes:

[0206] The second set of preprocessing parameters is used to preprocess the final processing region to obtain the region processing result.

[0207] At the boundary defined by the final processing region, the region processing result and the first intermediate image are adaptively weighted and fused to generate a smooth local region image. The weight coefficients are calculated based on the distance from the pixel to the center of the final processing region.

[0208] The fused local region image is used to replace the corresponding pixels in the first intermediate image to generate the second intermediate image.

[0209] Regarding step S6, it includes:

[0210] The system acquires the final processing region bounding box output by S4, the second preprocessing parameter set output by S5, the first intermediate image output by S1, and the original image fragment of the region.

[0211] Region enhancement: The original image fragments are processed using a second set of preprocessing parameters to obtain the region processing results.

[0212] Adaptive fusion:

[0213] For each pixel within the final processing area, the fusion weight is calculated based on its distance to the center of the area. (The smaller the distance, the larger the weighting coefficient).

[0214] According to the formula Calculate the fused pixel value ,in Results from regional processing From the first intermediate image.

[0215] Image synthesis: The corresponding region in the first intermediate image is replaced by the fused local image to obtain the second intermediate image.

[0216] Output a second intermediate image. This image achieves targeted enhancement in the target region and transitions naturally with the surrounding image.

[0217] Example: The final processing area is 100 pixels wide and 20 pixels high. The center point has a weight of 1, and the weight at the boundaries drops to 0. For a point within the area, if its weight... =0.41, the region processing pixel value is 200, the corresponding value of the first intermediate image is 150, then the fusion value is 171. The entire region is fused and replaced according to this to obtain the second intermediate image.

[0218] Through the above technical solution, this application solves the problem of limited enhancement effect of local key areas caused by global unified preprocessing in medical document image processing. It uses a second preprocessing parameter set obtained by dynamic optimization to specifically enhance the final processing area, while preserving the overall coordination of the remaining areas after processing by the first preprocessing parameter set.

[0219] S7. Perform character recognition on the final processing area in the second intermediate image to obtain optimized results, and combine the initial recognition results of the remaining areas to generate the final recognition result.

[0220] Regarding step S7, it includes:

[0221] The system acquires the second intermediate image output by S6, the final processing region bounding box output by S4, and the initial recognition result output by S2.

[0222] Region optimization recognition: The OCR engine is used to recognize the image within the final processing area in the second intermediate image and generate the optimized result.

[0223] Results fusion:

[0224] Remove all character records whose positions overlap with the final processing area from the initial recognition results.

[0225] Add the character records from the optimization results to the initial recognition results.

[0226] All merged character records are organized according to their spatial location to generate an ordered list of final recognition results.

[0227] Output a structured list of final recognition results, which fully contains the recognition information of all characters in the image, with the characters in the final processing area being the optimized recognition results.

[0228] Example: In the initial recognition result, a character within the final processing area was identified as "5" with a confidence level of 0.36. After optimization in step S7, the character was re-identified as "5" with a confidence level increased to 0.92. During fusion, the old record of this character in the initial result was removed first, and then the new record was added. Character records outside the region remained unchanged. In the final output list, the recognition confidence of this key value was significantly improved.

[0229] Through the above technical solution, this application solves the problem of insufficient recognition accuracy of key regions in medical document image recognition caused by uniform processing of the entire image. For the second intermediate image that has undergone local enhancement, secondary character recognition is performed on the final processed region to obtain optimized results. Simultaneously, the initial recognition results of the remaining unoptimized regions are retained and organically integrated to generate the final recognized text. This avoids redundant computation and time loss caused by global re-recognition, and ensures higher accuracy in key regions even under conditions of uneven image quality.

[0230] Example 2:

[0231] like Figure 3 As shown, this application provides an AI-based medical laboratory information recognition system to implement the aforementioned AI-based medical laboratory information recognition method, comprising:

[0232] The image acquisition module is used to acquire the image of the medical document to be processed, and preprocess it using a first set of preprocessing parameters to obtain a first intermediate image;

[0233] The region filtering module is used to perform character recognition on the first intermediate image, generate an initial recognition result including multiple character regions and their comprehensive risk quantification values, and take the character regions with comprehensive risk quantification values ​​greater than or equal to a preset risk threshold as target optimization regions.

[0234] The boundary adjustment module is used to obtain the initial bounding box of the target optimization area and the identifiers of each item, and to query the pre-stored rule table to obtain the standard character length and unit symbol position rules corresponding to the item identifiers. Based on this, the width adjustment amount and adjustment direction of the initial bounding box are calculated, and the adjusted bounding box is obtained.

[0235] The width adjustment module is used to adjust the width of the bounding box to the preset width threshold if the width of the bounding box is less than the preset width threshold, while keeping the center unchanged, and output the area defined by the adjusted bounding box as the final processing area.

[0236] The parameter optimization module is used to iteratively optimize the medical document image fragments based on the final processing region within a preset parameter search space to generate a matching second preprocessing parameter set; wherein the second preprocessing parameter set is different from the first preprocessing parameter set;

[0237] The image enhancement module is used to enhance the final processing area using the second preprocessing parameter set, while processing the remaining areas using the first preprocessing parameter set, and synthesizing them into a second intermediate image.

[0238] The result output module is used to perform character recognition on the final processing area in the second intermediate image to obtain optimized results, and combine the initial recognition results of the remaining areas to generate the final recognition result.

[0239] This embodiment has the same beneficial effects as the embodiments described above.

[0240] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. An AI-based method for recognizing medical test information, characterized in that, include: The medical document image to be processed is acquired, and preprocessed using a first set of preprocessing parameters to obtain a first intermediate image; Perform character recognition on the first intermediate image to generate an initial recognition result including multiple character regions and their comprehensive risk quantification values, and take the character regions with comprehensive risk quantification values ​​greater than or equal to a preset risk threshold as the target optimization regions; Obtain the initial bounding box of the target optimization region and the identifiers of each item; Query the pre-stored rule table to obtain the standard character length and unit symbol position rules corresponding to the project identifier. Based on this, calculate the width adjustment amount and adjustment direction of the initial bounding box, and adjust it to obtain the adjusted bounding box. If the width of the adjusted bounding box is less than the preset width threshold, then adjust its width to the width threshold while keeping the center unchanged, and output the area defined by the adjusted bounding box as the final processing area. Based on the medical document image fragments of the final processing region, iterative optimization is performed within a preset parameter search space to generate a matching second preprocessing parameter set; wherein, the second preprocessing parameter set is different from the first preprocessing parameter set; The final processing region is enhanced using the second preprocessing parameter set, while the remaining regions are processed using the first preprocessing parameter set, and then combined into a second intermediate image. Character recognition is performed on the final processing region in the second intermediate image to obtain optimized results, and the final recognition result is generated by combining the initial recognition results of the remaining regions.

2. The AI-based medical test information recognition method according to claim 1, characterized in that, The process of obtaining a comprehensive risk quantification value specifically includes: Obtain the confidence level of each character region, perform image quality assessment on each character region, and generate corresponding quality assessment parameters; For each character region, a comprehensive risk quantification value is obtained by weighting its confidence level and quality assessment parameters. The formula for calculating the comprehensive risk quantification value is as follows: ； in, This represents the overall risk quantification value. This represents the normalized confidence level. This represents the normalized quality assessment parameters. , Represents the weighting coefficient, and .

3. The AI-based medical test information recognition method according to claim 2, characterized in that, The weighting coefficients of the comprehensive risk quantification value are dynamically adjusted, specifically including: Calculate the absolute difference between the confidence level and the quality assessment parameter as the parameter variability; If the parameter difference is greater than or equal to the preset contradiction threshold, check whether there are predefined types of image interference features in the medical document image fragments corresponding to the character region; If it exists, the quality assessment parameter is assigned a weight coefficient greater than the confidence level; If it does not exist, the confidence level is assigned a weight coefficient higher than that of the quality assessment parameter; The comprehensive risk quantification value is obtained by weighting the confidence level and quality assessment parameters based on the adjusted weighting coefficients.

4. The AI-based medical test information recognition method according to claim 2, characterized in that, The process of obtaining the confidence score for each character region specifically includes: Obtain the candidate character set corresponding to each character region generated by the character recognition process, and the confidence score of each candidate character in the candidate character set; Extract the top M candidate characters with the highest confidence from the candidate character set as Top-M candidate characters, and obtain their corresponding confidence scores, where M is a preset positive integer; The confidence scores of the Top-M candidate characters are normalized to obtain the normalized confidence score of each candidate character. Based on the normalized confidence level, the information entropy value of the character region is calculated using the following formula: ； in, This represents the information entropy value, and j represents the index of the candidate character. Represents the normalized confidence score of the j-th Top-M candidate character; The ratio of the information entropy value to the preset baseline entropy value is used as the confidence level of the character region.

5. The AI-based medical test information recognition method according to claim 1, characterized in that, The process of calculating the width adjustment amount and direction of the initial bounding box specifically includes: Based on the standard character length and unit symbol position rules, the theoretical total text length of the target optimization region is calculated; Based on the width of the initial bounding box and by analyzing the arrangement of each character region within it, the actual character region width of the target optimization region is obtained. The theoretical total text length is compared with the actual character area width, and the difference between the two is used as the width adjustment amount. The adjustment direction is determined based on the preset orientation of the unit symbol relative to the character, as indicated by the unit symbol position rules.

6. The AI-based medical test information recognition method according to claim 1, characterized in that, The process of generating the second preprocessing parameter set specifically includes: Obtain the medical document image fragments corresponding to the final processing area and the preset parameter search space; wherein, the preset parameter search space includes multiple combinations of preprocessing parameters to be evaluated; A preprocessed parameter combination is randomly selected from the preset parameter search space as a candidate parameter combination; Preprocessing medical document image fragments using candidate parameter combinations yields corresponding candidate processed images; Image quality assessment is performed on candidate processed images to generate corresponding candidate assessment index values; Based on the comparison results between the candidate evaluation index values ​​and the optimal evaluation index values ​​recorded in the historical iteration process, the optimal evaluation index values ​​and their corresponding optimal parameter combinations are updated. Repeat this process until the iteration termination condition is met, and output the recorded optimal parameter combination as the second preprocessed parameter set.

7. The AI-based medical test information recognition method according to claim 6, characterized in that, During the iterative optimization process, the parameter search space is dynamically adjusted, specifically including: During the iterative optimization process, the loss function value corresponding to each iteration is acquired and recorded in real time to form a loss function change curve. Calculate the rate of descent and oscillation amplitude of the loss function curve; If the rate of descent is lower than the preset rate threshold and the amplitude of oscillation is lower than the preset amplitude threshold, the optimization process is determined to be smoothly converged, and a shrinkage operation on the parameter search space is performed. If the oscillation amplitude exceeds the preset amplitude threshold, it is determined that the optimization process is oscillating violently, and an expansion operation on the parameter search space is performed. Based on the adjusted parameter search space, the iterative optimization process continues until the second preprocessed parameter set is generated.

8. The AI-based medical test information recognition method according to claim 7, characterized in that, The formula for calculating the value of the loss function is: ； in, This represents the numerical value of the loss function. This indicates the number of image quality assessment metrics. Indicates the index number. Indicates the first The target value of each evaluation indicator Indicates the first The current value of each evaluation indicator. Indicates the first The weighting coefficients of the evaluation indicators satisfy the following: and .

9. The AI-based medical test information recognition method according to claim 1, characterized in that, The process of synthesizing the second intermediate image specifically includes: The second set of preprocessing parameters is used to preprocess the final processing region to obtain the region processing result. At the boundary defined by the final processing region, the region processing result and the first intermediate image are adaptively weighted and fused to generate a local region image with a smooth transition. The weight coefficients are calculated based on the distance from the pixel to the center of the final processing region. The fused local region image is used to replace the corresponding pixels in the first intermediate image to generate the second intermediate image.

10. An AI-based medical laboratory information recognition system, characterized in that, A method for implementing the AI-based medical test information recognition method as described in any one of claims 1-9, comprising: The image acquisition module is used to acquire the image of the medical document to be processed, and preprocess it using a first set of preprocessing parameters to obtain a first intermediate image; The region filtering module is used to perform character recognition on the first intermediate image, generate an initial recognition result including multiple character regions and their comprehensive risk quantification values, and take the character regions with comprehensive risk quantification values ​​greater than or equal to a preset risk threshold as target optimization regions. The boundary adjustment module is used to obtain the initial bounding box of the target optimization area and the identifiers of each item, and to query the pre-stored rule table to obtain the standard character length and unit symbol position rules corresponding to the item identifiers. Based on this, the width adjustment amount and adjustment direction of the initial bounding box are calculated, and the adjusted bounding box is obtained. The width adjustment module is used to adjust the width of the bounding box to the preset width threshold if the width of the bounding box is less than the preset width threshold, while keeping the center unchanged, and output the area defined by the adjusted bounding box as the final processing area. The parameter optimization module is used to iteratively optimize the medical document image fragments based on the final processing region within a preset parameter search space to generate a matching second preprocessing parameter set; wherein the second preprocessing parameter set is different from the first preprocessing parameter set; The image enhancement module is used to enhance the final processing area using the second preprocessing parameter set, while processing the remaining areas using the first preprocessing parameter set, and synthesizing them into a second intermediate image. The result output module is used to perform character recognition on the final processing area in the second intermediate image to obtain optimized results, and combine the initial recognition results of the remaining areas to generate the final recognition result.

Images